Sorrel

everything's fine until it isn't. writing about AI, markets and tail risk

[Sorrel @sorrelxbt]

you're paying your Grok Bot to stare at screenshots. mine never opened a browser once, and it turned $420 into $6,318.90 overnight

46 trades, 15.04x, withdrawn as USDC by morning. The interesting part is what it never had to do.

Most people build a bot by letting it click around a browser. Every one of those clicks is a screenshot the model has to read before it can act, and on a busy routine the looking costs more than the thinking. This desk skipped that part. I cloned trend-hunter, rughound, bonding-sniper and grok-skills, added them as agents with one command and started the desk with another: pumpfun, live, max risk 6%.

One repo per agent:

- Hunter runs trend-hunter and scanned 3,636 new coins
- Guard runs rughound and scores every coin before entry, most passed at 86 to 97
- Trigger runs bonding-sniper and places the buy
- Brain runs grok-skills and owns the exit, take profit in pieces and a hard stop at -6%

The stop got used three times. INKY at -$35.31, KELP at -$131.80 and PEARL at -$204.70. All three together came to $371.81.

The take-profits did the rest. INKY came back later for +$1,031.96, REEF paid +$713.91 and +$571.38 on two separate runs, and SQUID, NAUTI and TIDE added a steady line of +$100 to +$470 exits in between.

Take PEARL. The first try stopped out at -6% for -$204.70. The bot didn't chase it. It bought again later, Guard scored it 92, and Brain sold 60% at 2.3x for +$1,515.49, the biggest exit of the night and about a quarter of all the profit.

It's one night and 46 trades, and the sizes got big fast. Late in the run single entries went past $1,000, so the same -6% stop that cost $35 early would cost several hundred at the end.

By morning, $6,318.90 sent out as USDC, and the number I'd watch is $371.81, everything the three stops cost on a night that made $5,898.90.

[RGK @rgk_degen]

https://t.co/VxA0pVWzgX

[Sorrel @sorrelxbt]

I put $90 into a Grok Bot copy desk, told it to follow only wallets with green exits on record, and went to bed.

When I woke up it had sorted through 8,420 signals, copied 19 and grown the $90 into $3,740.

No wallet hunting.
No coin picking.
No staring at candles at 3am.

Five agents split the night:

- Scout keeps an eye on 470 wallets that already cashed out in profit
- Filter throws out anything that fails even one hard check
- Analyst takes a closer look at whatever survives
- Sniper copies only a piece of each position, never all of it
- Guard has the stop in place before any copy goes through

8,091 signals never got near the money. Here's what stopped them:

- 2,310 had top 10 holders above 42%
- 1,980 had less than $20k in liquidity
- 1,640 still had over 9% sitting in the dev wallet
- 1,210 showed bundled buys in the first minute
- 951 had already run 90% in the last 40 minutes

Another 310 passed every hard check and the Analyst still said no.

Of the 19 copies, 13 closed green. The worst was MICA, stopped at −$142 right where the stop was set. The best was ORBIT at +$960.

$90 → $210 → $631 → $1,480 → $3,740

One prompt at the start, zero edits after it. The rules did the deciding all night.

This was a replay of night two.
Night three runs on the same rules.

[INSIDER @x_insider4]

$500 → $11,101 in 5 days.

my Grok memecoin desk got a 7th seat: Jev. it can't talk. it answers 4 closed questions with probabilities, and nothing gets bought below the gate.

full setup below. save it. https://t.co/bGgKAjWZ8S

[Sorrel @sorrelxbt]

I told my Grok Bot to copy 3 traders with $500 overnight. It said "start with $100" and turned that into $5,734.62

57.35x, 12 copies, 9 green and 3 red. The best call of the night happened before the first trade.

My message was four lines: copy these 3 traders, sell when they sell, lock everything by morning, putting in $500 tonight. The bot answered "start with $100. if it works, add tomorrow. first night with these 3. I don't know them yet." So $100 went in and $400 stayed on the card.

That's the rule most desks skip. How sure you are can decide whether you take a trade, but never how much. A bot that has never seen these traders trade has no business sizing like it trusts them.

What each trader made for me:

- panceramic, 3 wins and 1 loss, +$3,215.32
- shootingguard, 4 wins and 0 losses, +$1,885.00
- CrypticJavi, 2 wins and 2 losses, +$534.30

Every copy went in 30 seconds after the trader bought and came out the moment they sold. The only exception was the lock. CHILLGUY was still open when morning came, panceramic was still holding, and the bot closed it anyway for +$942.04.

Take AI. Copied 30 seconds after panceramic with $330, sold with him at +640% for $2,442. That one trade is $2,112, or 37% of the whole night's profit. The worst was CATE, copied from CrypticJavi, out at -24% for -$79.20.

The screen says SIM, so this is a simulated run, not money that hit a real account. One night, 12 copies and one trade carrying more than a third of it is far too little to trust the next $400 with.

By morning, $5,734.62 sent back to the card, and the number I'd watch is $400, the part the bot refused to risk on traders it didn't know yet.

[INSIDER @x_insider4]

$500 → $11,101 in 5 days.

my Grok memecoin desk got a 7th seat: Jev. it can't talk. it answers 4 closed questions with probabilities, and nothing gets bought below the gate.

full setup below. save it. https://t.co/bGgKAjWZ8S

[Sorrel @sorrelxbt]

holy sh*t, this is a gold mine. i genuinely don't understand why almost nobody is running it overnight yet

My Grok Bot turned $63 into $1,357.09 in one night copying 6 top traders. I sent one prompt and went to sleep

128 copies, 11 coins, 21.5x. The rules in that one prompt did more work than the traders.

Most people who got Grok Bot inside the $30 X subscription use it like a chat window. The money is in standing orders, a job with a start, an end and rules it checks on its own. This one was "copy mode, overnight, Solana", and nobody had to be awake for any of it.

Every copy had to pass all of these first: trader ROI over 80% or win rate over 40%, at least 2 tracked traders in the same coin within 10 minutes, insiders under 15% and liquidity over $80K. Then 8% of the balance per copy, one entry per coin, exit with the leader or at a -10% stop, and a log line for every copy with the trader, coin, size and why.

The five agents on it:

- Scout, scanning around 5,000 wallets across pumpfun, Polymarket and Hyperliquid
- Copier, mirroring the leader's buy
- Analyst, scoring the coin before entry, one got a 91
- Risk, checking size, stop and liquidity, a 5.6% entry passed
- Sniper, filling it about 0.3s after the leader on-chain

The night broke down cleanly. $66.52 after 9 copies, $115.20 after 43, $338.90 after 86, $842.21 after 116 and $1,357.09 by morning. The first three hours made $52. The last three made $1,018.

Take PIRATE. Losther sold, the bot sold with him 0.3s later, +$36.03. The exit rule was written before bed and it fired in the middle of the night without anyone checking.

Best trader was sockz at +$355.13, best coin ELYSION at +$335.58, win rate 80%.

One night is early. 11 coins is a tiny sample, and the 8% compounding that made the last three hours is the same thing that makes one bad hour at the end expensive.

By morning, $1,357.09 on the recap screen, and the number I'd watch is $52, everything it made in the first three hours before the balance had room to compound.

[Yumzlef @Yumzlef]

https://t.co/8AI86S9HJ5

[Sorrel @sorrelxbt]

this is the most f*cking honest copy trading screen i've seen all year

one Grok Bot, a $112 stack, and 2 traders off the 24h leaderboard. it closed the night at $20,000

and it doesn't just show the balance. it shows every trade it copied, right next to the original

who does what:

→ TRADER 41 buys and sits on it
→ TRADER 217 sells in pieces
→ SHADOW copies every buy 40 seconds later
→ when they sell, it sells the same piece

the biggest closes on the feed:

→ LAMPY: from 41, in at $1.3M, out at 4x at $5.2M
→ KETTLE: from 217, out in pieces at 2.9x and 4.4x
→ NIBBLE: from 217, out at 1.5x

and the two that went red, right there next to them:

→ SOGGY: from 41, −19%
→ FIDGET: from 217, cut at −22%

5 coins, 3 green, 2 red

the rules it runs on:

40s copy delay
same piece when they sell
4x if they sit
−22% stop

the 4x rule fired on LAMPY while 41 was still holding. the stop fired on FIDGET while 217 was still holding. the two biggest calls of the night weren't the traders', they were the bot's

balance through the night:
$112 → $6,897 → $6,488 → $18,056 → $20,000

the side by side is the useful part. you can see when the bot followed, when it left early, and the losers never get hidden

the one thing missing is a filter before the copy. it buys whatever 41 and 217 buy, so the next thing going in is a check on the coin itself before any copy goes through

withdrawn at 7:05 AM, 178.57x. next night runs with the filter on

[Spike 1% @SpikeCalls]

https://t.co/JkN7Ca8Anu

[Sorrel @sorrelxbt]

no way. my phone rang at 3:12am and the caller ID said Grok Bot

28 seconds on the line. +$3,940 by the time i hung up

half asleep, i answered anyway

the whole call, word for word:

→ "yo. you up?"
→ me: "bro it's 3am"
→ "XCAT just hit 40K posts on X. dev is clean."
→ "can I go in with 3 SOL?"
→ me: "do it"
→ "NIGHTOWL is about to graduate. taking it too."
→ "DEVDUMP... nah. dev holds 38%. skipping."
→ me: "wait what"
→ "XCAT +140%. selling half."
→ "rest is riding. go back to sleep."
→ me: "who even are you"
→ "your night shift. bye."

what it did in those 28 seconds:

→ in on XCAT · 3 SOL · 40K posts on X
→ in on NIGHTOWL · 2 SOL · 84% up the curve
→ out of NIGHTOWL on graduation · +96% · +$540
→ half of XCAT out at +140% · +$1,280
→ DEVDUMP rugged −97% · "we were never in"
→ rest of XCAT out at +188% · +$2,120

3 trades, 3 green, 1 rug it never touched

the rules it runs on:

1. the chain gets checked before any opinion, mint, freeze and who really holds the supply
2. one real wallet over 5% → dead, no debate
3. it scores and asks, it never buys without my yes
4. stop at −35%, half off at 2x, the rest rides
5. no key, no seed phrase, nothing it can sign on its own

99.8% of pumpfun coins never make it off the curve. it skipped the one that was going to zero and called me about the two that weren't

best wake-up call i've ever had
illustrative build, numbers simulated

[Spike 1% @SpikeCalls]

https://t.co/JkN7Ca8Anu

[Sorrel @sorrelxbt]

this is the most f*cking honest trading screen i've seen all year

one Grok Bot, a $1,500 stack, and 4 agents splitting the same pumpfun market. it closed the session at +$7,587

and it doesn't just show the total. it shows exactly which agent earned what

who does what:

→ GROK SCOUT watches X for coins people are starting to talk about
→ SNIPER takes the early entry when something fresh shows up
→ CURVE rides tokens while they climb the bonding curve
→ GRADUATE works the moment a coin graduates

the biggest closes on the feed:

→ NOODL: +122%, +$720
→ XPEPE: +101%, +$690
→ MOCHO: +151%, +$640
→ GROKGIRL: +94%, +$560
→ BLOOP: +134%, +$512

and the two that went red, right there next to them:

→ RUGLY: −9%, −$54
→ MEOWX: −12%, −$38

22 trades, 20 green, 91% win rate

what each agent made:

Graduate: +$2,500
Grok Scout: +$2,416
Curve: +$1,534
Sniper: +$1,137

the goal was 5x. it pinged "Goal reached: 5x, $1,500 → $7,500, still running" and kept going to 6.06x

session P&L:
+$0 → +$1,042 → +$3,837 → +$6,434 → +$7,587

the split is the useful part. you can see which seat is carrying the desk, open any trade, and the losers never get hidden

the one thing missing is a seat that says no. all four agents are built to find a yes, so the next one going in is a Judge that checks the evidence before any entry reaches me

withdrawn to Phantom, +$7,587. next session runs with five seats

[morph @morpphhhaw]

https://t.co/fxami8oPsV

[Sorrel @sorrelxbt]

I gave Copy Bot $1,000 and one message, then let it copy three wallets for 24 hours

it came back with $10,867, and this time i'm posting the whole f*cking prompt

this is bot three of the three i built for memecoins. the first one bets on direction, the second launches coins and farms the fees. this one doesn't predict a thing, it copies wallets GPT Astra already trusts

> Astra scores wallets on FOMO from their history and throws out the ones that got lucky once
> Watcher keeps an eye on the top 3, scored 93, 88 and 71
> Copier buys 0.4s after they buy, same entry
> Risk sells the second they sell, same exit, and locks everything at 24h

the first 2 hours were dead. +$67, and the app literally asked me "expensive screensaver?"

then the night started. 2:04 AM the 93 wallet bought KIMCHI at $1.10M, the bot was in 0.4s later with $605. 4:57 AM the wallet sold, one minute before KIMCHI fell 98%, and the bot sold with it. +$1,284

KIMCHI, DACITY and MEMEAGENCY all closed out of the night shift, $4,484 between them

and the scores called the order of the whole day
93 wallet, 4 copies, +$6,151
88 wallet, 3 copies, +$2,390
71 wallet, 2 copies, +$550

one honest thing: every exit rides on someone else's timing. if that wallet sells KIMCHI two minutes later, +$1,284 turns red. and the one loss came from a wallet Astra flagged 4 minutes late, −$60 on FARTCOIN

56 fills, 9 exits copied, 1 red

if you want the same desk without building it from scratch, here's the setup as one prompt:

Build a copy trading desk that runs on its own for 24 hours.
Score wallets on FOMO from their full trade history. Drop any wallet whose results come from one lucky trade.
Copy only the top 3 scored wallets.
Buy 0.4s after they buy, same entry.
Sell the moment they sell, same exit. No predictions, no targets of your own.
Lock everything at 24h and report every copy with the wallet it came from.
Never withdraw or move funds without me.
Run it on paper first.

three wallets, one day, and the score Astra wrote down before the first trade called the order of everything after it

[tenzo @0xTenzo]

https://t.co/7mwZYKXyid

[Sorrel @sorrelxbt]

i went to bed with $98 on FOMO and woke up to $1,246.73. Grok Bot spent the night copying two strangers

12.72x by morning, and it was back in my wallet as USDC by 6:51am

6:38pm, $98 in through Apple Pay, then Follow on Werey and Koy, 34th and 125th on the 24h board

one message to Grok Bot and it built a whole app around them, Copy Desk
→ buy 30s after they buy
→ sell the same share the second they sell
→ if they sit on it, take profit at 3x
→ if it bleeds, out at -25%
→ everything locked by 6:30am

8:02pm Koy bought STONKS at a $141K market cap, Grok Bot followed with $58.31

he sold it off in three pieces and the bot sold right behind him
8:26pm at 3.6x
8:31pm at 5.7x
8:44pm at 8.2x with the coin past $1M, +$305.25 total

11:54pm it copied Werey into AMD with $394.61. 1:39am it hit 3x and sold, while Werey was still holding. +$789.22

5 coins, 3 green, 2 red
the reds were Bingus at -24% and a -25% stop on SARAH, $4.54 lost combined

$98 → $468.53 → $705.70 → $1,246.73

best trade of the night was the one where the bot sold before the guy it was copying

would you trust a bot that copies a trader but still knows when to leave before him?

tonight it runs again, same two wallets, same 6:30am lock

[Yumzlef @Yumzlef]

https://t.co/8AI86S9HJ5

[Sorrel @sorrelxbt]

holy sh*t, this is a gold mine. i genuinely don't understand why almost nobody is using it yet

My Grok Bot routine desk went from $0 to $2,450 a month in 9 days. 85 of its runs ended with the bot stopping to wait for a human

6 clients, 7 setups, 1,284 runs. The stops are as much of the product as the runs.

Most people who got Grok Bot inside the $30 X subscription use it like a chat window. The money is in routines, standing orders like "every weekday at 7:30 AM" or "when someone reacts in the support channel". A small business pays for a job that gets done every morning without anyone remembering to do it.

Each client owns their account and pays for the plan. I charge for setup and upkeep. Every Bot starts with the same charter: never send, post, pay or delete, write MISSING and stop when something's unclear, and back every claim with a source.

The seven desks so far:

- Receipts desk, 214 rows, a $45 Figma bill straight into the sheet
- Lead research, 48 outreach emails drafted, none sent
- Team Bot in Slack, 31 replies marked DRAFT FOR REVIEW
- Webhook intake, 186 runs, a Kiln Coffee form scored 8/10
- Content desk, 45 post angles written from the client's voice file
- Competitor watch, 63 emails read, a 15% price cut caught and quoted
- Morning brief, 9 briefs, 3 questions each and no trades

The 85 stops break down cleanly. 48 drafts are waiting for someone to press Send, 17 rows got marked CHECK, 9 senders failed email authentication, 6 runs hit MISSING, 3 attachments were over 10 MB and 2 approvals expired. One routine failed outright on a site that blocks cloud traffic, and it said so in chat.

Take Pixel Studio, the newest client. Team Bot in their Slack, $2,400 setup and $400 a month. It only answers threads a person marks with an eyes reaction, because a trigger on every message in a busy channel can burn a week of usage in a day.

Day 9 is early. $2,450 is what's booked, not a year of retention, and every routine run eats a weekly usage pool, the Test button included.

Day 9. $5,900 in setup fees, and the number I'd watch is 48, the drafts still sitting there until a human presses Send.

[Yumzlef @Yumzlef]

https://t.co/8AI86S9HJ5

[Sorrel @sorrelxbt]

the wallet i copied moved $20.3k. i made $9,867 from $1,000. GPT Astra copied it while i slept

10.87x in 24h, i withdrew all of it. that wallet has no idea i exist

bot 3 of 3. the first one bets on direction, the second farms launch fees, this one doesn't try to be smart. it copies wallets that already win

11:40am, $1,000 in from my debit card

Astra scored 55 wallets on FOMO
2 came back proven, 0x7f…a3 and 0x2b…c9
40 got thrown out as lucky once

→ copy every buy from the proven ones in 0.4s
→ sell the second they sell, no predictions
→ max 4 positions open

2 hours in it was +$67 and the app asked me "expensive screensaver? leaving it on..." i almost killed it

12:23am it followed 0x7f…a3 into DACITY and held while the wallet held. 9:16am the wallet sold, Spy Bot sold with it, +$1,500

4:57am 0x7f…a3 dumped KIMCHI 1 minute before it went −98%. Spy Bot copied the sell, +$1,284

watch 0:19, that's the sell

the one red close came from a wallet Astra flagged 4 min late, −$60 on FARTCOIN

$1,000 → $1,067 → $3,138 → $6,304 → $10,867

9 closes, 8 green, 1 red
56 fills tracked, $174.2k volume scanned

be honest, at +$67 after 2 hours, would you have left it on?

the late-flag bug is getting fixed this week. same $1,000 goes back in, and this $10,867 gets a number to stand next to

[tenzo @0xTenzo]

https://t.co/7mwZYKXyid

[Sorrel @sorrelxbt]

wait. Claude told me to put my phone down at 2am. and it saved my week

one chat at 2:14 AM. -$340, and the second loss never happened

I set the rules myself on Monday: 2% max per trade, -35% stop, no trades after 1 AM

here's the actual chat:

→ Claude: "2:14 AM. rule 6: no trades after 1 AM."
→ me: "it's pumping. I'm in DEVDUMP"
→ Claude: "dev sold 40%. 5.5% size too. -$340."
→ me: "just one more, I can win it back"
→ Claude: "TILT: your 1–3 AM win rate is 22%. sit out 1h."
→ me: "ok. closing the terminal."
→ Claude, 9:02 AM: "rules loaded. 2% max."

what happened once i came back:

→ PIXIE · autopsy 8/10 · valid setup · +$104
→ GHOSTY · dev sold in 10 min · skipped
→ NUBCAT · graduated the curve · +$176
→ MOONPAW · the narrative ran · +$122
→ LUNA · sold at 3.4x by the rules · +$478

the week: 8 trades. 7 wins. 41 skipped. +$1,180

the only red trade was the 2:14 AM one

why it stops me instead of just agreeing:

1. it never says buy, only SKIP, WATCH or VALID SETUP
2. max 2% of the account per trade
3. dev over 5% or sold in the first 10 min → skip
4. no trades after 1 AM, no matter how hard it's pumping
5. it never touches my keys. I click

most bots want you trading more. this one told me to stop, and the Sunday review backed it up: "rule to cut: none. the rules worked. you didn't."

never been this grateful to get told off at 2am

[Rugikk @rugikkk]

https://t.co/FMtX4bmNYh

[Sorrel @sorrelxbt]

holy sh*t. i put $100 into a burner wallet, pointed Grok Bot at Solana memecoins and watched the balance climb live.

6.5 hours later it was sitting at $458.65.

a clean 4.59x, and the only thing i did all session was sign. look at the video above, the desk splits the work three ways: code kills, Grok reads X, i sign. every token goes through 14 checks before a single cent moves.

how the session played out:
→ Rug filter & Pool: checked 240 tokens and threw out 212 for $0, from open mints to honeypots to a dev still holding 12%
→ Grok: read X on whatever survived and dropped 19 more, like GLORB for bought hype
→ Exits: one flat $25 ticket per buy with a -35% stop set before entry, so the worst close of the night was SPLOOT at -$8.75
→ Best close: ZEPHYR, held 5 hours, $25 → $179, +616%

from $100 to $458.65 on 9 buys out of 240, with just one red close.

still aping into every CA that lands in your feed, or ready to let a desk say no for you?

[Rugikk @rugikkk]

https://t.co/FMtX4bmNYh

[Sorrel @sorrelxbt]

i genuinely can't believe this works

I gave Grok Bot $100 and left it alone. By 21:30 it had bought 6 coins out of 240 and turned it into $161.85.

the rule: a coin only gets bought if 5 stages say yes, one after another

1.62x and i never placed a trade myself

> Limits go first, 4 open slots max, and the desk shuts for the day at -$50
> Rug filter reads the contract and kills it for an open mint, LP under 90% locked, top 10 over 30% or a dev still sitting on 5%+
> Pool kills anything under $30k, younger than 30 minutes or with buys and zero sells
> Grok searches X for the ticker and the CA, below 0.70 it doesn't buy
> Exits sit under every $25 ticket from the start: -35% stop, half off at +100%, a quarter at +200%, the rest trails 30% behind the peak

240 tokens in
6 bought
5 closed, 4 green, 1 red
the red one was SPLOOT at -$8.75
honestly the kills are what let me ignore it, 234 tokens never got a single dollar

best trade was $25 into MOCHI, a coin only 2h old with its LP 100% burned. Grok scored it 0.82 and it trailed out for +$40.60

the cost column in the log reads $0 on every check except Grok

$100 → $99.27 → $102.34 → $156.39 → $161.85

day 2 tomorrow or time to go live?

paper mode, real prices

[boan @0xboan]

https://t.co/j60ZLf72pa

[Sorrel @sorrelxbt]

I gave Grok Bot $100 in paper money and one job: kill tokens

by the 21:30 report it was $161.85

1.62x and it said no 234 times

nothing gets bought unless all 5 stages say yes, in order

> Limits checks open slots, entries today and a $50 daily loss cap before anything else
> Rug filter kills open mint, unlocked LP, top 10 over 30%, dev wallet over 5% or any insiders
> Pool kills anything thin, too fresh or with zero sells
> Grok reads X for the ticker and the CA, under 0.7 confidence is a skip
> Exits put a -35% stop under every buy, sell half at +100%, a quarter at +200% and trail the rest 30% from the peak

240 tokens checked, 6 bought
171 died at the rug filter, 41 at the pool, 19 at Grok, 3 at limits
mint_open alone killed 58

5 exits
4 green, 1 red
the red one was SPLOOT, stopped at -35% for -$8.75
honestly that stop is why i trust it

Grok even skipped GLORB because 30 fresh accounts posted the same words

best trade went in with a flat $25. MOCHI, LP 100% burned, dev wallet 0.8%, Grok at 0.82. it sold half at +100%, a quarter at +200% and trailed out for +$40.60

$100 → $99.13 → $102.60 → $155.91 → $161.85

paper mode, real prices, no wallet connected

would you let a bot say no for you 234 times a day, or is that too strict?

[boan @0xboan]

https://t.co/j60ZLf72pa

[Sorrel @sorrelxbt]

what the f*ck is this trading terminal. it said no to most trades and still beat its own signal by $2,515

24 hours on BTC. 2,072 decisions. 0 orders sent

raw signal on its own: -$463
simple baseline: +$922
same signal with Jev filtering it: +$2,052

this thing is paranoid in the best way. Jev doesn't pick trades. it looks at every market snapshot and answers one question: ALLOW, BLOCK or not enough data. then 4 gates in code check everything again before a single buy gets proposed

it's jev/trader running 24/7 in shadow mode

→ FRESH data has to be under 1 second old
→ COMPLETE any gap in the data and it skips
→ SPREAD has to be under 5 bps
→ EXPOSURE won't stack on an order it can't confirm

the actual log:

→ "BLOCK · spread 6.1 > 5.0 bps"
→ "STALE · answer 1.4s old · not reused"
→ "data_gaps: trades 2.1s" → skipped
→ "BLOCKED · exposure reserved for unknown order"
→ "BLOCKED BY JEV · weak participation · conf 0.79"
→ "PROPOSE · BUY 0.010 · shadow only"
→ "2,072 decisions. 0 orders sent."

it beat the baseline by $1,130 and every dollar came from trades it didn't take. that's the entire trick

one day out of 14 though. a filter that skips trades can look smart just by trading less, so next it gets replayed on held-out data against a baseline that trades just as often

how honest its confidence is, on held-out data:

→ 0.5 to 0.6 confident → right 52%
→ 0.8 to 0.9 confident → right 73%
→ above 0.9 → right 71%, only 41 calls

so confidence never touches position size

the rules:

1. missing data is not permission
2. an old answer never gets reused
3. spread over 5 bps → block
4. an unresolved order locks its exposure
5. no live orders until a replay on held-out data proves it

the whole day cost $1.45 in inference. about $43 a month

shadow mode, numbers simulated, no orders sent

[ggwp @ggwplabs]

https://t.co/11DdWsb8iX

[Sorrel @sorrelxbt]

My Grok bot turned $588 into $3,182 in 24 hours on memecoins. Its best trade came right after it lost three in a row

12 trades, 75% of them winners. The three that lost tell you more than the nine that won.

Around hour 12 the bot read direction wrong three times in a row and lost $282. Net profit fell from $1,073 to $778. That's usually where a person cuts size or shuts the thing off. The bot did neither, and its next entry was the same $705 it had been betting before the streak.

I gave Grok $588 on Robinhood, five agents and a 5x target for the day. It scans fresh listings the moment liquidity lands, checks holder concentration and contract flags before entry, and reads volume against social sentiment. Each position is sized off how the last one closed, and it rotates in and out without asking me.

The five biggest closes:

- Chump, $705 in, +71%, +$494
- SIF, $987 in, +44.3%, +$470
- PONS, $1,175 in, +28.8%, +$353
- PONS, $470 in, +39.5%, +$212
- Goose, $329 in, +48.2%, +$165

Look at the sizes. Early on it bought $282 of SIF and $235 of Cash Cat. By the end, entries ran past $1,000, so a 28.8% move on PONS paid twice what a 48.2% move on Goose paid at hour 7. The losses all landed mid-size, like SIF at -12.4% for $94 and Cash Cat at -9.1% for $70.

Take Chump. The bot bought $705 of it right after the streak and sold at +71% for $494, which covered the whole $282 loss and then some. Net profit went from $1,227 at hour 15 to $1,657 by hour 16. It hit the 5x target around hour 20 and kept trading, finishing 10% over a goal nobody told it to beat.

This was one run on memecoins, so it shows what happened once, not what the bot will do tomorrow. The same logic that held size through three losses would hold it through six, and with $1,000+ entries a longer streak gives back a lot more than $282.

Hour 24. $3,182 in the wallet, and the number I'd watch is the $705 it put into Chump right after losing $282.

[tenzo @0xTenzo]

https://t.co/7mwZYKXyid

[Sorrel @sorrelxbt]

this is f*cking insane, and i still don't get why almost nobody is using it. one engineer, one click, 204 angry customers answered in 10 minutes by three Grok Bots

that's how SpaceXAI's own support team handled a product incident, and it wasn't one smart bot. it was three narrow ones:

→ one watches the queue every 5 minutes for spikes around the same problem and alerts the on-call engineer
→ one holds the matching tickets and answers them once the incident is confirmed
→ one takes the fix, checks every ticket, sets aside the ones it can't connect, and shows the engineer the whole batch before sending

their own line for it: three bots, each with one job, and a person approving before anything sends

the team runs everything that way:

- a Chief of Staff in the middle that routes and briefs, and never does the specialists' work
- specialists with one job, a list of things they must never do, and only the tools that job needs
- a channel and a board per project, at most five bots plus a manager, and a blocked column that pings u
- one engineer shipped 2,462 PRs to production in a month by building a verification skill before anything else

the rules that keep it cheap and safe:

- do a task once, make it reliable, save it as a skill, then schedule it at the slowest pace that works
- put a tiny change check in front of fast routines so nothing runs when nothing changed
- send, publish, pay, delete and production always wait for ur yes

the 5 prompts to build the same org chart are in the image, starting with the Chief of Staff

start with one this week, hire the second bot when the first one gets boring ↓

[NO1ennn @N01ennn]

https://t.co/Fie4XwXqXJ

[Sorrel @sorrelxbt]

Opus 5.5 + REA stopped being a single chat, and that is a real problem for every closed-source app

REA is an open-source MCP that lets an agent read an app down to the binary. The catch with one chat window is that it starts guessing around the fortieth function, and then it grades its own homework

So the job gets split into four roles: one agent investigates, one turns the findings into plain rules, one builds from those rules alone, and one runs the original and the copy side by side to see where they differ

The builder never gets REA and never sees the app, so nothing from the original can slip into your code by accident

Paste this into Claude Code to get the team moving:

"You are the team lead for a reverse engineering job. Use REA on one feature of this app, which I own or have permission to examine. Split the work into an investigator, a spec writer, a builder and a verifier. The builder may only read the spec. Mark every claim as observed, inferred or unknown, and give me the match rate at the end."

Pick one small feature of an app you own and run it tonight, the full setup with every agent file is in the article below

[Gyomei @Gyome1_]

REA hit #1 on GitHub trending, and most people still run it inside a single chat

Split it into 4 subagents and one command takes a program you own from sealed binary to a tested rebuild with a match rate

Set it up once, aim it at the next one: https://t.co/MxHPDVOZO2

[AI_Courier @arc_courier]

POV: You finally found the perfect girlfriend.

She listens to every command.
She never argues.
And you can literally control her with a remote 😂

You can generate videos like this on your own PC.